WhitneyLand

There was some improvement in cognitive scores, but no placebo group. Without a placebo group, there are a lot of explanations for the data.

It might be saturated for smaller scopes of work, but it’s not hard to see the cracks when you scale up what you ask of SOTA models/agents.

One example, to try and single shot prompt coding a ChatGPT equivalent chatbot.

Sure it will spit something out, but the feature depth, UX subtitles, backend integration, and lots of pragmatic engineering decisions along the way will just not be baked.

Another example is building a C compiler from scratch which Anthropic showed is still a struggle to do.

Not that these these specific examples are important but just to point out scaling up expectations shows the cracks.

It’s not just a model problem of course, better agents, orchestration features (like Dynamic Workflows mentioned in the post), all need to continue to evolve.

Ar what point does my CS degree become totally useless is an open question.

The big difference is that meditation and journaling do not require a belief that you are communicating with supernatural beings.

“I don't think intelligence and spiritual practice are mutually exclusive.”

That’s a low bar. At the least we know supernatural/religious beliefs are negatively correlated with scientific training and scientific eminence.

This is the kind of design wisdom that’s both true and difficult to win an argument over.

It reminds me of arguments related to over-engineering and complexity. The principles are super important to having a codebase that scales and continues to be efficient to work in as the team grows, but they are hard to objectively measure.

Locally or in isolation something may sound like a great idea. Being able to step back and see the greater ripple effects require some experience and intuition that can’t always be used to convince people otherwise.

The problem is you can’t defend it right? Someone could say your evidence came from a prompt: “Take this article and reverse engineer a hypothetical unpolished first draft written in a mix of Russian and English”

I’m not sure what the right answer is here. Fwiw I have no doubt you wrote it unassisted.

The other day I was criticized for posting a comment people thought was AI but was actually not.

I’m starting to notice that more often with others as well. Happens sometimes to those who were always using emdash, sometimes to those who happen to have traits that these machines themselves learned from how to write, and now they sound suspicious.

I don’t think this means we should never call out slop or lazy writing, but it does seem our ability to detect this stuff is on a spectrum. Some of it is obvious. But beyond a certain point, for example with this article, the signals can become too weak to make any strong claims.

It’s disconcerting to admit that we’ve come to a point where it’s possible to be completely fooled one way or the other by what’s human or AI. Lots of stuff we can still detect, and sometimes it’s obvious, but at the margins we can no longer reliably discriminate.

One could argue coding is only a use case and that their models are still killing it overall. However agents are strategic across the board and coding agents are at the forefront. They’ve already lead to new products like CoWork and it’s easy to understand why Google should be doing everything possible to catch up.

Surprised they’re not trying to entice developers away with more heavily subsidized subscription plans. Maybe it’s because as some say those days are ending and soon we’ll all be paying per token. Or maybe it would just put too much of a strain on available compute.

The point is you made two brigade style comments about my posts sounding suspiciously like an LLM and having hallucinations.

Neither turned out to be true and I think a better response would concede the point.

It may be more helpful for us to stick together as humans since we can’t always recognize each other so easily anymore.

1. Evidently you’re no longer able to distinguish AI from people as the whole comment was written by a human off the cuff.

2. The numbers are not hallucinations. It’s word on the street reporting, so yes it’s speculative, but a model did not make up it up unless that’s where TechCrunch got it which is not on me.

https://techcrunch.com/2026/04/17/sources-cursor-in-talks-to...

Ffs.

edit: removed cursing you out. Sorry but this is frustrating. I don’t leave AI generated comments here (or anywhere else).

Fwiw it’s beating Claude Sonnet in most benchmarking (benchmaxxing?), yet they’ve priced it almost half off on a per token basis.

Question is are you going to persuade anyone with this argument?

Are there many devs at Google who legit prefer Gemini over Claude and Codex? Would love to hear about that.

Forking VS Code, going big on bleeding edge features like cloud agents, and now they’ve thrown down the gauntlet directly challenging frontier labs by training their own model (“much larger” than Kimi 2.5’s 1T parameters) from scratch.

They’ve been highly successful so far. Raised $50B, $2B in revenue, forecast to end 2026 above $6B. But even at these heights, they’re just not in the same league as OpenAI/Anthropic/Google.

And if building a state of the art multitrillion parameter model is not challenging enough, it’s a mountain you don’t climb just once. Every few months you need to push it farther with a new release. Fall off for a couple cycles and like Facebook you may never catch up again.

Not for the faint of heart.

Would interesting to see the success rate for Claude Cowork or Codex’s equivalent feature.

Not really.

- The benchmarks are based on F8_E4M3 and you’re not running that on any Mac.

- Sonnet has a 1M token context window. This is 256k but again you’re probably not even getting that locally.

- Sonnet is fast over the wire. This is going to be much slower.

“Productivity soars when a computer and its users interact at a pace (<400ms)”

If that momentum is strongest with Lean so be it. At the same time things become more machine verifiable, converting to a new system will also become easier. It can already be strongly assisted using a general agent like Claude Code.

“Peter Thiel and Marc Andreessen have parlayed their extensive ties with the president into an unabashed assault on universities and institutional science. In private text messages leaked to The Washington Post last year, Andreessen wrote that “universities are at Ground Zero of the counterattack.” He characterized Stanford and MIT as “mainly political lobbying operations fighting American innovation at this point” and vowed that universities would “pay the price” after “they declared war on 70% of the country.” Most troublingly, Andreessen called for the National Science Foundation to receive “the bureaucratic death penalty.””

It doesn’t mean profitable, it doesn’t mean cash flow positive.

Are you just trying to say their revenue is greater than zero?

It’s not just Zed, CoPilot also reduces the capabilities and options available when using models directly.

No thanks, definitely agree with the Open Router approach or native harness to keep full functionality.

You’re making a claim, and I’m pointing out that it’s unsubstantiated and not consistent with any other source of data, including that internal to the company that makes the model.

I hope you can see that that’s different than saying it’s worked well for me

What wild claim to make. Unsupported by benchmarks, unsupported by the consensus of the community, no evidence provided.

Sounds like in another comment here even the GLM5 team concedes they are behind the frontier wrt tool calling, do you know something they don’t?

Bright White Lightning, but didn’t see a track name.

Overall, wow.

Example 2TB:

Google $10/mo vs S3 ~&45/mo?

You could get cheaper that Google Drive with glacier tiers but that’s a different level of restrictions and still has retrieval fees.

The idea is explained by Rick Beato here: https://youtu.be/rGremoYVMPc